A method for target tracking and aiming of a UAV

Abstract

The present application provides a kind of unmanned aerial vehicle target follow sight method, this method is with " offline foreground distillation training-online light detection-target special time-frequency memory association-bounded freezing-two-stage relocking-follow sight deviation output" as main line, offline training stage, in training stage, construct teacher network and light student network, based on the training image of the unmanned aerial vehicle to be followed, the multiscale features of teacher network and student network are extracted, multiscale foreground response distillation loss is calculated based on the multiscale features of teacher network, to train student detector with light detection head.Online follow sight stage, light student detector frame by frame outputs unmanned aerial vehicle candidate set, candidate association energy is used to select the most consistent candidate with historical target, update memory when candidate is reliable, when candidate is unreliable, bounded freezing is executed, short loss adopts local search, continuous loss is switched to global search, and relocking is completed through candidate interval and continuous multiple frame confirmation.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, specifically to a method for tracking and aiming at targets on unmanned aerial vehicles (UAVs). Background Technology

[0002] UAV laser energy transfer transmits energy to the receiver of the flying UAV via a directional laser beam. To ensure the laser beam acts stably on the receiving area, the vision front end needs to continuously output the position of the designated UAV and the target center. When multiple UAVs with similar appearances appear simultaneously at long distances, on small scales, with strong background interference, the vision system must not only detect "whether a UAV exists" but also maintain the continuity of identity by "whether the current target is still the originally designated UAV".

[0003] Existing general-purpose UAV detection methods typically rely on larger backbone networks and higher-dimensional detectors to improve accuracy. While this enhances the representation of small targets, it increases the number of parameters and computational load. Directly using lightweight backbones, on the other hand, can easily result in the loss of edges, contours, and local responses of UAVs at long distances. Conventional knowledge distillation methods often impose the same constraints on all feature regions, which weakens the effective supervision of the foreground of small targets when the background region accounts for too large a proportion.

[0004] On the other hand, using the highest confidence detection box as the tracking result frame by frame is affected by detection box fluctuations, missed detections, and competition from similar targets, causing the target center to jump. Traditional tracking methods based on intersection-union ratio or linear motion only utilize a single geometric quantity, making it difficult to handle scale changes, short-term occlusion, and interference from similar candidates simultaneously; continuous unbounded velocity extrapolation after target loss may also produce abnormal drift, causing the output position to deviate rapidly from the true target.

[0005] Therefore, a complete algorithm is needed that is adapted to the laser-based tracking and aiming requirements of drones. Summary of the Invention

[0006] This invention aims to solve the following problems: lightweight detectors are insufficient for characterizing small-scale UAVs; it is difficult to maintain the identity of a specified target among similar UAV candidates based solely on the confidence of a single frame; frame-by-frame detection box jitter affects the stability of laser pointing; unbounded extrapolation after short-term target loss causes abnormal drift; and there is a lack of a unified relocking mechanism for local occlusion recovery and long-term loss recovery.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for tracking and aiming unmanned aerial vehicles (UAVs) includes the following steps: S1, acquiring training images and video frames containing the UAV to be tracked; S2, during the training phase, constructing a teacher network and a lightweight student network, extracting multi-scale features of the teacher network and student network based on the training images of the UAV to be tracked, calculating multi-scale foreground response distillation loss based on the multi-scale features of the teacher network, and training a student detector with a lightweight detection head; S3, during the continuous tracking and aiming phase, inputting the current video frame into the trained student detector to obtain a UAV candidate set; S4, establishing a target-specific memory based on the locked targets, wherein the target-specific memory models the target center, scale, velocity, confidence, and frequency domain stability; S5, for each candidate target in the UAV candidate set, calculating the candidate association energy composed of position consistency, bounding box overlap consistency, scale consistency, trajectory frequency domain stability, and detection confidence, and determining the target locking state based on the minimum association energy, candidate interval, and reliability threshold; S6, when the candidate target meets the locking conditions, outputting the target bounding box and target center. Step S7: When a candidate target does not meet the locking condition, perform bounded freezing that limits the prediction speed and the number of consecutive prediction frames. Step S8: After bounded freezing, perform local candidate search and full-map candidate search based on the prediction center in sequence, sort the relocking candidates by associated energy and confirm them for multiple consecutive frames. After successful relocking, restore memory update and output the two-dimensional tracking deviation of the target center relative to the optical axis reference point.

[0008] Preferably, in step S2, the lightweight student network includes a PREsNet18 lightweight backbone network, a detection decoding head, and an efficient multi-scale context compensation module; the teacher network is only used to generate distillation supervision during the training phase and does not participate in inference during the continuous tracking phase.

[0009] Preferably, in step S4, the target-specific memory includes at least a short-term location sequence, a short-term scale sequence, a smoothing velocity, a long-term target feature prototype, historical detection confidence, and frequency domain stability obtained from the discrete Fourier transform of the target center increment sequence.

[0010] Preferably, in step S5, the correlation energy of the j-th candidate target is obtained by a weighted combination of normalized position distance, crossover ratio cost, logarithmic scale change cost, trajectory high-frequency energy ratio and detection confidence, and the candidate target with the smallest correlation energy is selected as the first candidate, where j is an integer not less than 1.

[0011] Preferably, in step S6, the target locking conditions include at least the association energy of the first candidate being less than a preset acceptance threshold, and the difference in association energy between the first candidate and the second candidate being greater than a preset identity interval threshold.

[0012] Preferably, the bounded freezing in step S7 includes: calculating the exponential smoothing velocity based on the historical target center, pruning the upper limit of the velocity amplitude, and limiting the number of lost frames to a preset maximum number of freezing steps, so as to generate a bounded prediction center and a bounded prediction box.

[0013] Preferably, in step S8, the radius of the local candidate search region increases adaptively with the number of lost frames and the speed of historical targets; when the local candidate search does not meet the relocking condition within a preset number of frames, it switches to the full-map candidate search.

[0014] Preferably, the relocking conditions include at least the relocking candidate association energy being lower than the relocking threshold, the first relocking candidate and the second relocking candidate satisfying the identity interval, and the same candidate satisfying the conditions in K consecutive frames, where K is an integer not less than 1.

[0015] This invention follows a main framework of "offline foreground distillation training - online lightweight detection - target-specific time-frequency memory association - bounded freezing - two-stage relocking - tracking deviation output". In the offline training phase, a high-capacity teacher detection network and a lightweight student detection network extract multi-scale features in parallel. An expanded foreground mask is constructed based on target annotations, highlighting the target region with teacher feature responses as weights, while applying weak consistency constraints to non-target regions. After training, the teacher network and distillation branch are deleted, retaining only the lightweight student detector. In the online tracking phase, the lightweight student detector outputs a UAV candidate set frame by frame; the target-specific memory models the target center, scale, velocity, confidence, and frequency domain stability; candidate association energy is used to select the candidate most consistent with historical targets. The memory is updated when a candidate is reliable, and writing stops and bounded freezing is performed when a candidate is unreliable; short-term loss uses local search, while continuous loss switches to global search, and relocking is completed through candidate intervals and confirmation over multiple consecutive frames.

[0016] This invention achieves the following technical effects: By extending foreground supervision, teacher response weighting, weak background consistency, and multi-scale weighting, the lightweight student detector primarily inherits the teacher network's sensitive representation of small-scale UAVs, without increasing the inference branches of the student model after training. The lightweight student detector, through the Lite192 detector head and efficient multi-scale context compensation, maintains small target detection performance while reducing the number of detector parameters and computational cost. By jointly correlating energy in the temporal, scale, and frequency domains, it simultaneously incorporates target motion continuity, bounding box scale changes, and high-frequency trajectory jitter into candidate selection, reducing target switching caused by similar candidates. Bounded freezing limits motion speed and continuous extrapolation duration prevents rapid target box drift during the missed detection phase; a two-stage search from local to global and continuous multi-frame confirmation improves the reliability of re-locking. The output target center's two-dimensional deviation relative to the optical axis can be used as visual input for UAV laser power transfer, free-space optical link acquisition, and other designated target tracking systems. Attached Figure Description

[0017] Figure 1 This invention provides an offline training method for a UAV target tracking and aiming method.

[0018] Figure 2 This invention provides an online continuous tracking and aiming method for UAV targets.

[0019] Figure 3 The diagram shows the training architecture of the multi-scale foreground response distillation detector provided by this invention.

[0020] Figure 4 The relationship diagram between candidate correlation energy and gating memory update provided by this invention.

[0021] Figure 5 A flowchart of bounded time-frequency memory tracking and relocking state provided for the invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1

[0024] Figure 1 The present invention provides an offline training method for a UAV target tracking and aiming method, such as... Figure 1 As shown, during the offline training phase, multi-scale features are extracted in parallel by a high-capacity teacher detection network and a lightweight student detection network 101. An expanded foreground mask is constructed based on the target annotation 102. The target region is highlighted with the teacher feature response as the weight, and a weak consistency constraint is applied to the non-target region 103. After training, the teacher network and distillation branch are deleted, and only the lightweight student detector is retained, thus obtaining the lightweight student detector 104.

[0025] Figure 2 The present invention provides an online continuous target tracking method for unmanned aerial vehicles (UAVs), such as... Figure 2 As shown, during the online tracking phase, the lightweight student detector outputs a candidate set of UAVs frame by frame. The target-specific memory models the target center, scale, velocity, confidence, and frequency domain stability. The candidate association energy is used to select the candidate that is most consistent with the historical target. When the candidate is reliable, the memory is updated. When the candidate is unreliable, writing stops and bounded freezing is performed. For short-term loss, local search is used, and for continuous loss, global search is switched. Relocking is completed through candidate intervals and confirmation of multiple consecutive frames.

[0026] Specifically, during the continuous tracking phase, the current video frame is input into the trained student detector 201 to obtain a candidate set of UAVs. Then, a target-specific memory 202 is established based on the locked targets. The target-specific memory models the target center, scale, velocity, confidence, and frequency domain stability. For each candidate target in the UAV candidate set, the candidate association energy 203, which is composed of position consistency, bounding box overlap consistency, scale consistency, trajectory frequency domain stability, and detection confidence, is calculated. The target locking state 204 is determined based on the minimum association energy, candidate interval, and reliability threshold. When a candidate target meets the locking conditions, the target is locked normally 205, and the target bounding box and target center are output 208. When a candidate target does not meet the locking conditions, a bounded freeze is performed to limit the prediction speed and the number of consecutive prediction frames 206. After the bounded freeze, a local candidate search and a full-map candidate search are performed sequentially based on the prediction center. The re-locked candidates are sorted by associated energy and confirmed for multiple consecutive frames 207. After successful re-locking, the memory is restored and updated, and the two-dimensional tracking deviation of the target center relative to the optical axis reference point is output 208.

[0027] Figure 3 The training architecture diagram of the multi-scale foreground response distillation detector provided by this invention is as follows: Figure 3 As shown, training images and annotations (301) are first generated. Then, the annotated images are input into the teacher network detector and student network detector to extract multi-scale features of the teacher branch (302) and student branch (303). Subsequently, based on the extracted teacher branch multi-scale features (302), two processing steps are performed: foreground region expansion (304) and teacher response weight generation (305). Based on the student branch multi-scale features (303), weak background consistency (306) constraint processing is performed. The student multi-scale features, optimized by the above constraints, are input into the Lite detection head (308) to complete the training of the lightweight detection module. Finally, a smaller detection model with faster inference speed is obtained. After all training steps converge iteratively, the entire training process is completed, resulting in a deployable lightweight detection model. For the l-th layer teacher features... and student characteristics First, use channel projection Student features are mapped to a dimension consistent with teacher features. A binary foreground mask is generated at the feature layer based on the target bounding boxes. And obtain the expanded foreground mask through scale-dependent dilatation. : (1) In equation (1), and Let the width and height of the target be the feature width at layer l. The radius of expansion is determined by the foreground. Smaller targets correspond to a larger relative expansion range to complement edge and neighborhood responses.

[0028] Teacher response weights are generated based on the channel average absolute response of teacher characteristics. : (2) The expanded foreground feature loss, teacher response weighted loss, and weak background consistency loss are respectively: (3) (4) (5) The total loss of multi-scale foreground response distillation is: (6) In equation (6), To normalize the scale weights, Weights used to reduce background consistency loss. The overall training objective of the student network consists of detection classification loss, bounding box regression loss, intersection-union loss, and... composition.

[0029] In this embodiment, the student detector uses PResNet18 as its backbone, compresses the hidden dimension of the detection decoder head from 256 to 192, and sets up an efficient multi-scale context compensation module before the multi-scale features enter the decoder. For input features... Local mapping and global pooling mapping are performed separately to obtain the compensated features: (7) in, and GAP represents lightweight linear or convolutional mapping, and GAP represents global average pooling. This represents local pooling at scale k. This module supplements different scale contexts without changing the form of the detection output.

[0030] In this embodiment, the candidate extraction and target-specific memory process is as follows: Current frame Candidate set obtained by lightweight student detector . No. The candidates are determined by the bounding box. ,center Confidence level and target features Composition. Target-specific memory. This includes the target center, scale, smoothed velocity, long-term target prototype, and reliability records from the most recent L-frame. The target center velocity is smoothed exponentially. (8) Figure 4 The candidate correlation energy and gating memory update relationship graph provided by this invention, such as Figure 4 As shown, this is a weighted energy evaluation and adaptive gating update process for target tracking scenarios. Multi-dimensional input is collected, including candidate bounding boxes and positional accuracy (401), positional distance (402), overlap cost (403), scale cost (404), and frequency domain jitter (405). Based on the collected multi-dimensional indicators, a comprehensive energy calculation is performed using a preset weighted combination formula (406). Combining the weighted energy, candidate bounding box positional accuracy, and inter-frame interval information from the previous step, the reliability of the current tracking state (407) is calculated, quantifying the credibility of the current tracking result. An exponential moving average method is used to adaptively update the gating parameters (408). When the current tracking reliability is deemed insufficient or the tracking state unreliable, the system stops the regular state writing process and switches to a bounded freeze / recapture mechanism to prevent the tracker from drifting erroneously in abnormal scenarios such as target occlusion or loss, waiting for the target to reappear before resuming tracking (409). (Based on memory prediction center) and prediction boxes Calculate the location cost, intersection-union ratio cost, and scale cost of the candidates respectively: (9) (10) (11) The candidate centers are added to the center sequence of the most recent L frames, and the discrete Fourier transform of the center increment sequence is calculated. The proportion of high-frequency energy to total energy is used as the frequency domain jitter cost caused by the candidates. (12) Candidate correlation energy is defined as: (13) Where wp, wi, ws, wf, and wc are the weight coefficients for each dimension, and the weighted summation yields an energy value that comprehensively reflects the current tracking state. The candidate with the lowest energy is selected. And calculate the second candidate The energy gap between the first candidate and the first candidate. The first candidate satisfies... and When the condition is met, the system enters a stable locked state; when only some conditions are met, it enters an uncertain state; otherwise, it enters a bounded frozen state.

[0031] In this embodiment, the controllability-gated memory update process is as follows: the candidate reliability is calculated based on the detection confidence, candidate energy, and candidate interval: (14) when When the update threshold is reached, a gated update rate is used. Update the long-term goal prototype and its state of motion: (15) (16) Reduce when the candidate is in an uncertain state When entering bounded freeze and recapture states, command This prevents false detections or other similar drone characteristics from being written into the long-term target prototype.

[0032] Figure 5 The bounded time-frequency memory tracking and relocking state flowchart provided for the invention is as follows: Figure 5 As shown, the detection candidate set Dt of the current frame and the historical memory Mt-1 retained from the previous frame 501 are used to comprehensively calculate the association energy between the current candidate box and the historical target based on the candidate box and its position degree, position distance, overlap cost, scale cost, and frequency domain jitter 502, quantifying the matching degree between the current candidate and the historical target. The best candidate selected is double-checked 503 to check whether the association energy reaches the preset threshold and whether the inter-frame interval meets the tracking continuity requirements. When the candidate is determined to be reliable, the target's position, motion speed, and appearance prototype are directly updated to maintain a stable tracking state 504. When the state is uncertain, the write rate of the memory module is reduced 505 to avoid erroneous tracking information from polluting the historical memory. When the candidate is determined to be unreliable, the process enters the bounded freezing stage, which performs bounded pruning on the velocity prediction range and prediction duration to limit the extrapolation range of the tracker and prevent the tracking box from drifting irregularly after the target is lost 506. When the target is completely lost, the system initiates a hierarchical recapture strategy, prioritizing local recovery (507): performing a small-scale search within the predicted target center neighborhood to quickly retrieve the target. If local recovery fails, global recovery (508) is triggered, reordering all detection candidates across the entire map and rematching the lost target. The total output shows the target center coordinates deviating from the optical axis (509). Strong constraints are also set, allowing historical memory updates only when the tracking status is reliable, thus preventing erroneous information from interfering with the stability of long-term tracking.

[0033] When the target is lost, the smoothing velocity is truncated in magnitude, and the number of consecutive prediction steps is limited: (17) (18) In the formula, For the maximum allowed inter-frame rate, This is the maximum number of freeze steps. This prevents the target box from being pushed out indefinitely during periods of missed detection.

[0034] During the local recovery phase, a radius is constructed with the bounded prediction center as the center. Search area: (19) If local recovery fails within a preset number of frames, switch to global search. In addition to meeting the associated energy threshold and candidate interval, relocking candidates must also remain consistent across K consecutive frames. (20) Finally, use an image or a calibrated optical axis reference point ( , Based on ), the center of the target bounding box is ( , If the target is 2D, then the aiming deviation will be output as follows: (twenty one) The output data also includes the lock status, target bounding box, candidate reliability, and number of lost frames. The deviation can be used as a visual observation for subsequent beam pointing control, but this invention is not limited to a specific servo mechanism, computing platform, or controller.

[0035] The experimental verification process of this invention is as follows: a detection experiment was conducted on the DUT-Anti-UAV dataset with an input size of 640×640. Representative results are shown in the table below.

[0036]

[0037] Compared to RT-DETRv2-R18, the detector of this invention reduces the number of parameters from 20.084M to 15.768M, FLOPs from 30.692G to 23.056G, while improving AP from 0.6380 to 0.6464 and APS from 0.553 to 0.575. Under the same lightweight structure, the average AP of the three random seeds—without distillation, with the original DGBD, and with the foreground response distillation of this invention—are 0.6362±0.0026, 0.6406±0.0012, and 0.6471±0.0008, respectively, indicating that the training mechanism simultaneously improves accuracy and training stability.

[0038] On the DUT-Anti-UAV tracking sequence, the same lightweight detector output was used as the input to the detection-driven tracker, and the comparison results are as follows. AUC, AvgIoU, P20, and NP0.5 measure the localization accuracy, Lost is the number of frames with an inter-cross-union ratio of 0, and Jitter is the measure of target center trajectory jitter.

[0039]

[0040] The AUC of the method in this invention is close to that of ByteTrack-lite, but the jitter is reduced from 6.7248 to 4.7558, showing a more suitable accuracy-stability balance for continuous laser tracking. The unconstrained motion extrapolation of SORT-lite leads to a jitter of 169.3621, which also experimentally illustrates the necessity of imposing boundary constraints on loss prediction.

[0041] When using only frame-by-frame detection, the jitter was 23.9920, which decreased to 6.3003 after introducing basic memory. After adding bounded prediction, trajectory jitter was further reduced while the AUC remained basically unchanged. Finally, after adding two-stage correlation, the Lost decreased to 404 and the jitter decreased to 4.7558. It can be seen that memory is used to suppress frame-by-frame fluctuations, temporal-scale-frequency domain correlation is used to improve candidate consistency, bounded prediction is used to limit anomaly extrapolation, and two-stage search is used to improve relock stability.

[0042]

[0043] To verify the ability of the source domain-trained model to transfer to another UAV imaging domain, the DUT-trained model was fine-tuned in the target domain on the Anti-UAV-RGBT visible light training set. The AP (Average Power) of the DUT model when tested directly in the target domain was 0.395; the AP was 0.532 when trained in the target domain using only a general backbone initialization without DUT detector pre-training; and the AP was 0.542 when fine-tuned in the target domain after DUT detector pre-training. These results demonstrate that source domain detection knowledge can provide positive transfer for target domain adaptation. This experiment illustrates the algorithm's domain adaptation capability without altering the DUT dataset's positioning as the primary detection and tracking validation domain of this invention.

[0044] This invention follows a main framework of "offline foreground distillation training - online lightweight detection - target-specific time-frequency memory association - bounded freezing - two-stage relocking - tracking deviation output". In the offline training phase, a high-capacity teacher detection network and a lightweight student detection network extract multi-scale features in parallel. An expanded foreground mask is constructed based on target annotations, highlighting the target region with teacher feature responses as weights, while applying weak consistency constraints to non-target regions. After training, the teacher network and distillation branch are deleted, retaining only the lightweight student detector. In the online tracking phase, the lightweight student detector outputs a UAV candidate set frame by frame. The target-specific memory models the target center, scale, velocity, confidence, and frequency domain stability. Candidate association energy is used to select the candidate most consistent with historical targets. The memory is updated when a candidate is reliable, and writing stops and bounded freezing is performed when a candidate is unreliable. Short-term loss uses local search, while continuous loss switches to global search, and relocking is completed through candidate intervals and confirmation over multiple consecutive frames.

[0045] The present invention achieves the following technical effects: (1) By expanding foreground supervision, teacher response weighting, weak background consistency and multi-scale weighting, the lightweight student detector mainly inherits the teacher network's sensitive representation of small-scale UAVs, and does not increase the inference branch of the student model after training. (2) By using the Lite192 detector head and efficient multi-scale context compensation, the small target detection performance is maintained while reducing the number of detector parameters and computation. (3) By jointly associating energy in the temporal, scale and frequency domains, the continuity of target motion, changes in box scale and high-frequency jitter of trajectory are simultaneously included in the candidate selection, reducing the target switching caused by similar candidates. (4) By using reliability-gated updates, long-term target prototype writing is stopped in the low confidence, occlusion or candidate conflict stage to avoid erroneous candidates polluting the memory. (5) By limiting the motion speed and continuous extrapolation duration through bounded freezing, the target box drifts rapidly in the missed detection stage; the reliability of re-locking is improved through two-stage search from local to global and continuous multi-frame confirmation. (6) Output the two-dimensional deviation of the target center relative to the optical axis, which can be used as the visual input for UAV laser power transmission, free space optical link capture and other designated target tracking systems.

[0046] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0047] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0049] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for tracking and aiming at a target on a drone, characterized in that, Includes the following steps: S1. Acquire training images and video frames containing the drone to be tracked; S2. During the training phase, a teacher network and a lightweight student network are constructed. Multi-scale features of the teacher network and the student network are extracted based on the training images of the UAV to be tracked. Multi-scale foreground response distillation loss is calculated based on the multi-scale features of the teacher network to train the student detector with a lightweight detection head. S3. During the continuous tracking phase, the current video frame is input into the trained student detector to obtain the drone candidate set; S4. Establish a target-specific memory based on the locked target, wherein the target-specific memory models the target center, scale, velocity, confidence level and frequency domain stability; S5. For each candidate target in the UAV candidate set, calculate the candidate association energy, which is composed of position consistency, bounding box overlap consistency, scale consistency, trajectory frequency domain stability and detection confidence. Determine the target locking status based on the minimum association energy, candidate interval and reliability threshold. S6. When a candidate target meets the locking condition, output the target bounding box and the target center.

2. The UAV target tracking and aiming method according to claim 1, characterized in that, It also includes step S7, which involves performing bounded freezing that limits the prediction speed and the number of consecutive prediction frames when the candidate target does not meet the locking conditions.

3. The underwater small target tracking method according to claim 2, characterized in that, It also includes step S8, which involves sequentially performing local candidate search and full-map candidate search based on the prediction center after bounded freezing, sorting the relocked candidates by associated energy and confirming them for multiple consecutive frames, restoring memory update after successful relocking, and outputting the two-dimensional tracking deviation of the target center relative to the optical axis reference point.

4. The UAV target tracking and aiming method according to claim 1, characterized in that, In step S2, the lightweight student network includes a PREsNet18 lightweight backbone network, a detection decoding head, and an efficient multi-scale context compensation module; the teacher network is only used to generate distillation supervision during the training phase and does not participate in inference during the continuous tracking phase.

5. The UAV target tracking and aiming method according to claim 1, characterized in that, In step S4, the target-specific memory includes at least a short-term location sequence, a short-term scale sequence, a smoothing velocity, a long-term target feature prototype, historical detection confidence, and frequency domain stability obtained from the discrete Fourier transform of the target center increment sequence.

6. The UAV target tracking and aiming method according to claim 1, characterized in that, In step S5, the correlation energy of the j-th candidate target is obtained by a weighted combination of normalized position distance, crossover ratio cost, logarithmic scale change cost, trajectory high-frequency energy ratio and detection confidence, and the candidate target with the smallest correlation energy is selected as the first candidate, where j is an integer not less than 1.

7. The UAV target tracking and aiming method according to claim 6, characterized in that, In step S6, the target locking conditions include at least the association energy of the first candidate being less than a preset acceptance threshold, and the difference in association energy between the first candidate and the second candidate being greater than a preset identity interval threshold.

8. The UAV target tracking and aiming method according to claim 2, characterized in that, The bounded freezing in step S7 includes: calculating the exponential smoothing velocity based on the historical target center, pruning the upper limit of the velocity amplitude, and limiting the number of lost frames to a preset maximum number of freezing steps, so as to generate a bounded prediction center and a bounded prediction box.

9. The method according to claim 3, characterized in that, In step S8, the radius of the local candidate search region increases adaptively with the number of lost frames and the speed of historical targets; when the local candidate search does not meet the relocking condition within a preset number of frames, it switches to the full-map candidate search.

10. The method according to claim 1 or 9, characterized in that, The relocking conditions include at least the relocking candidate association energy being lower than the relocking threshold, the first relocking candidate and the second relocking candidate satisfying the identity interval, and the same candidate satisfying the conditions in K consecutive frames, where K is an integer not less than 1.

Images